Skip to content

Commands

The model-cards plugin exposes a single slash command, /model-card, with two subcommands today and three more on the roadmap.

/model-card create

/model-card create <model-name> [--provider X] [--out path]

Researches a single model and produces a Mitchell-extended card after a human-in-the-loop review checkpoint.

Argument Required Description
<model-name> yes Model name as published by the provider, e.g. claude-opus-4-7, gpt-5-mini, meta-llama/llama-4.
--provider X no Provider hint — e.g. anthropic, openai, google, meta, mistral, xai, cohere, alibaba. If omitted, the agent infers from the model name.
--out path no Library directory override. Cards still land beneath this path as <provider>/<model-name>.md. Highest-priority path resolver.

Path resolution

The resolved target path is <library-root>/<provider>/<model-name>.md, where <library-root> is the first of:

  1. --out flag value (highest priority)
  2. MODEL_CARDS_DIR environment variable
  3. ~/.claude/model-cards/ (default)

Flow

  1. Parse args — model name, provider hint, output path.
  2. Resolve target path using the priority order above.
  3. Existing-card prompt — if the target exists, ask overwrite / skip / load-existing-as-base.
  4. Dispatch the model-card-researcher agent. Output is either a markdown card body or a REFUSED: string.
  5. Handle REFUSED — if the agent's output starts with REFUSED:, surface the refusal verbatim and abort. No file is written.
  6. Show review summary — sources used per section, sections that came up thin, top 3 most-cited claims with URLs, estimated research token cost.
  7. Disposition prompt — accept / edit / re-run-section <N> / abort.
  8. Validation checkpoint (on accept) — frontmatter parseable, all 10 sections present in canonical order, every [T<n>.<m>] citation resolves to a frontmatter source, Section 10 fields use field-level "Not publicly available" rather than silent omission. Deviations are fixed in place; agent is not re-dispatched.
  9. Write the card to the resolved target path. Print Card written: <full-path>.

See the research a model card how-to for a step-by-step walkthrough.

Disposition options

Choice Effect
accept Validate, then write the card to the target path.
edit Open the draft in $EDITOR (or vi if unset); on save, the edited content replaces the draft and the disposition prompt re-asks.
re-run-section <N> Re-dispatch the agent with a section-specific prompt for template section N (1-10). Replaces just that section in the draft, then re-prompts.
abort Discard the draft. No file is written.

Exit behaviours

  • REFUSED — model existence not confirmed via tier-1 + tier-2. Refusal reason printed; no file written. Exits with status 0.
  • Abort — user chose abort or the existing-card prompt returned skip. No file written. Exits with status 0.
  • Validation failure — citation references a missing source index in frontmatter. Surfaced to user before write; user is given the chance to edit or abort. No file written until validation passes.
  • Success — card written; full path printed. Exits with status 0.

/model-card seed

/model-card seed [--force]

Bulk-populates the library with cards for a shipped list of 14 frontier models.

Argument Required Description
--force no Overwrite existing cards in the library. Without --force, existing cards are skipped.

Flow

  1. Read the seed list at ${CLAUDE_PLUGIN_ROOT}/seed/frontier-models.json — format [{"name": "...", "provider": "..."}, ...].
  2. Print the list and total count.
  3. Single confirmation — Research <N> cards into <library-root>? [y/N]. Anything other than y aborts.
  4. For each model (sequential):
  5. Resolve target path (same logic as create).
  6. If card exists at target path AND --force is not set, skip with a one-line message.
  7. Dispatch model-card-researcher.
  8. If REFUSED, log the skip with reason (one line); continue.
  9. Else, write the card and log creation (one line).
  10. Print summary:
Created: N
Skipped (existed): N
Skipped (refused): N
Failed (other): N

Idempotency

seed is idempotent for the common-case "ran partially, want to retry" pattern — the existence check at step 4 skips already-written cards. To re-research existing cards, either pass --force (overwrites all) or run /model-card create <name> per-model (interactive review).

See the seed your library tutorial for an end-to-end walkthrough.

Roadmap subcommands

The following subcommands are tracked as GitHub issues but are not yet implemented:

Subcommand Issue Purpose
/model-card list #233 Browse the library — list cards by provider, by date researched, by version.
/model-card compare <a> <b> #234 Side-by-side card comparison across the 10-section structure.
/model-card refresh <name> #235 Re-research a specific card non-interactively, preserving the existing card as the base.

These are deliberately deferred — create + seed cover the bootstrap flow for v0.1.0. The shape of list, compare, and refresh will firm up as adoption signal arrives.